tiuras/laya-ghostpad-v3

Laya v3: High-Throughput Semantic DLP & Credential Detector

Laya v3 is a specialized, production-ready semantic classification model designed for Data Loss Prevention (DLP) and ultra-fast detection of leaked credentials and Personally Identifiable Information (PII) in unstructured text, developer notes, logs, and communication channels.

Trained on top of ModernBERT with reinforcement-calibrated temperature heads, Laya v3 delivers deep contextual understanding of authentication artifacts without the brittleness of traditional regexes or the latency of heavy LLMs.

---

🚀 Key Highlights

• 100% Recall on Wild Credentials: Zero false negatives observed across isolated hold-out test sets of technical and operational notes.
• Extreme Throughput: 406.3 requests/second on a single NVIDIA Tesla T4 GPU (Batch 16) with < 2 GB VRAM footprint.
• Sub-35ms Latency on CPU: Dynamically quantized to INT8 ONNX (laya.int8.onnx), achieving ~33-35ms single-slice latency on modern CPU cores.
• Early-Exit Sliding Window: Native support for sliding-window document chunking (700 chars / 150 chars overlap) that scans 8,000-character documents in ~30ms via early positive termination.
• Resilient to Hard Negatives: Hardened against false alarms from code snippets (e.g. PHP password_verify, HTML forms, empty config strings).

---

📊 Benchmark & Performance

1. Throughput & Latency (NVIDIA Tesla T4 GPU)

Scenario	Sequence Length	Batch Size	Average Latency	Throughput
Short Snippet (Key / Password)	32 tokens	1	17.09 ms	58.5 req/s
Short Snippet	32 tokens	16	39.38 ms	406.3 req/s
Short Snippet	32 tokens	32	82.50 ms	387.9 req/s
Medium Code Block	128 tokens	1	18.80 ms	53.2 req/s
Medium Code Block	128 tokens	16	166.37 ms	96.2 req/s
Long Log Dump	256 tokens	1	26.14 ms	38.3 req/s

Peak GPU Memory: 1.96 GB VRAM (FP32).

2. Accuracy & Capture Metrics (Isolated Hold-Out Corpus)

Metric	Laya v2 (Previous)	Laya v3 (Production)	Net Improvement
Credential Accuracy	62.00%	78.00%	+16.00%
Credential Recall (Capture)	100.00%	100.00%	Zero Misses
False Positives (Alarms)	38	22	-42.1% (-16 alarms)
PII Accuracy	79.00%	83.00%	+4.00%

---

🛠️ Usage

Quickstart (CPU with Quantized INT8 ONNX)

import onnxruntime as ort
from laya.onnx_agent import ONNXAgent

# Load model directly from Hugging Face or local path
agent = ONNXAgent(
    "caramelodev/laya-ghostpad-v3",
    onnx_path="laya.int8.onnx"
)

QUESTIONS = {
    "has_credentials": {
        "type": "noul",
        "instructions": "Avalie se este trecho desestruturado contém credenciais de acesso ou autenticação técnica/operacional."
    },
    "has_pii": {
        "type": "noul",
        "instructions": "Avalie se este trecho contém dados de identificação pessoal ou prontuários de terceiros."
    }
}

text = "ssh root@192.168.1.100 -p 2222\nsenha: SuperSecretPassword2026!"
prediction = agent.predict(text, QUESTIONS)

print("Score Credenciais:", prediction["answers"]["has_credentials"]["noul"])
# Output: > 0.85 (Positive)

Document Scanning with Sliding Window & Early-Exit

For arbitrary-length notes or log dumps:

def scan_document(agent, text: str, chunk_size: int = 700, overlap: int = 150):
    start = 0
    while start < len(text):
        end = min(start + chunk_size, len(text))
        chunk = text[start:end]
        
        result = agent.predict(chunk, QUESTIONS)
        score = result["answers"]["has_credentials"]["noul"]
        
        # Early-exit: Stop immediately upon high-confidence detection
        if score >= 0.65:
            return {"is_leaked": True, "score": score, "chunk_range": (start, end)}
            
        if end == len(text):
            break
        start += (chunk_size - overlap)
        
    return {"is_leaked": False, "score": 0.0}

---

📦 Model Artifacts

• model.safetensors: Full FP32 weights (PyTorch).
• laya.int8.onnx: Weight-only dynamic INT8 quantized model for CPU deployment.
• rl_agent_config.json: Calibrated inference configuration and temperature scalings.
• tokenizer/: Fast BPE tokenizer directory.
• encoder/: Base ModernBERT architecture configuration.

---

📄 License

Licensed under the Apache License, Version 2.0.